Project-Oriented Scheduler for Cluster Systems

Abstract

Parallel processing is the key to fulfill the high demands on computational resources in scientific computing. This has further pushed research in High Performance Computing into the mainstream. Numerous powerful computer systems have appeared. Especially, low-cost powerful clusters, which are set up by connecting many personal computers/workstations via a high speed network, have been developed rapidly during the last decade. Batch scheduling systems for clusters are very important on transparent access to cluster resources. Most batch scheduling systems have essentially focused on maximizing the use of computing resources like processors, but not on improving quality of services (QoS). This paper presents a batch scheduler called Project-Oriented Scheduler (POS), which schedules jobs from projects with different priorities. The higher priority level one project has, the longer service time it is assigned. Moreover, starvation is also considered. POS has been evaluated using SimGrid, a simulation tool that provides core functions for the simulation of distributed applications in distributed environments, and the results show that POS improves not only the utilization of the system but also the satisfaction of the projects as compared with other scheduling strategies.